{"id":26246,"date":"2025-05-11T15:48:39","date_gmt":"2025-05-11T07:48:39","guid":{"rendered":"http:\/\/139.9.1.231\/?p=26246"},"modified":"2025-05-11T15:48:40","modified_gmt":"2025-05-11T07:48:40","slug":"asr-en-parakeet-tdt-0-6b-v2","status":"publish","type":"post","link":"http:\/\/139.9.1.231\/index.php\/2025\/05\/11\/asr-en-parakeet-tdt-0-6b-v2\/","title":{"rendered":"\u82f1\u6587\u8bed\u97f3\u8bc6\u522b\u6a21\u578b\uff1aParakeet TDT 0.6B V2"},"content":{"rendered":"\n<p class=\"has-text-align-center\"><a href=\"https:\/\/huggingface.co\/nvidia\/parakeet-tdt-1.1b\"><strong><em>https:\/\/huggingface.co\/nvidia\/parakeet-tdt-1.1b<\/em><\/strong><\/a><\/p>\n\n\n\n<p><code>parakeet-tdt-1.1b<\/code>\u00a0\u662f\u4e00\u4e2a\u81ea\u52a8\u8bed\u97f3\u8bc6\u522b (ASR) \u6a21\u578b\uff0c\u53ef\u5c06\u8bed\u97f3\u8f6c\u5f55\u4e3a\u5c0f\u5199\u82f1\u6587\u5b57\u6bcd\u3002\u8be5\u6a21\u578b\u7531\u00a0<a href=\"https:\/\/github.com\/NVIDIA\/NeMo\">NVIDIA NeMo<\/a>\u00a0\u548c\u00a0<a href=\"https:\/\/www.suno.ai\/\">Suno.ai<\/a>\u00a0\u56e2\u961f\u8054\u5408\u5f00\u53d1\u3002\u5b83\u662f FastConformer [1] TDT [2]\uff08\u7ea6 11 \u4ebf\u4e2a\u53c2\u6570\uff09\u6a21\u578b\u7684 XXL \u7248\u672c\u3002<\/p>\n\n\n\n<p>\u82f1\u4f1f\u8fbe\u5728\u53d1\u5e03\u4e86\u4e00\u6b3e\u5f00\u6e90\u8bed\u97f3\u8bc6\u522b\u6a21\u578b\uff1a<strong>Parakeet TDT 0.6B V2<\/strong>\uff0c\u5176\u4ee5 600M \u53c2\u6570\u767b\u9876 Hugging Face Open ASR \u699c\u5355\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image size-full\"><img loading=\"lazy\" width=\"910\" height=\"415\" src=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/05\/image-21.png\" alt=\"\" class=\"wp-image-26252\" srcset=\"http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/05\/image-21.png 910w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/05\/image-21-300x137.png 300w, http:\/\/139.9.1.231\/wp-content\/uploads\/2025\/05\/image-21-768x350.png 768w\" sizes=\"(max-width: 910px) 100vw, 910px\" \/><\/figure>\n\n\n\n<p>\u5e73\u5747\u8bcd\u9519\u8bef\u7387\uff08WER\uff09\u4ec5 6.05%\uff0c\u8d85\u8d8a\u6240\u6709\u4e3b\u6d41\u95ed\u6e90\u6a21\u578b\u3002\u5b83\u80fd\u5728 1 \u79d2\u5185\u8f6c\u5f55 60 \u5206\u949f\u9ad8\u8d28\u91cf\u97f3\u9891\u3002<\/p>\n\n\n\n<p>\u57fa\u4e8e FastConformer \u67b6\u6784\u548c TDT \u89e3\u7801\u5668\uff0c\u4ec5\u7528 600M \u53c2\u6570\u5b9e\u73b0\u8d85\u4f4e WER \u548c\u6781\u5feb\u63a8\u7406\u901f\u5ea6\u3002\u8be5\u6a21\u578b\u57fa\u4e8e NVIDIA NeMo \u548c Suno \u56e2\u961f\u6536\u96c6\u548c\u51c6\u5907\u7684 64K \u5c0f\u65f6\u82f1\u8bed\u8bed\u97f3\u8fdb\u884c\u8bad\u7ec3\u3002<\/p>\n\n\n\n<p>\u8be5\u6a21\u578b\u91c7\u7528 <strong>FastConformer-TDT \u67b6\u6784<\/strong>\u3002<\/p>\n\n\n\n<p><strong>FastConformer<\/strong> \u662f\u5bf9\u4f20\u7edf Conformer \u6a21\u578b\u7684\u4f18\u5316\u7248\u672c\uff0c\u91c7\u7528\u4e86 <strong>8 \u500d\u6df1\u5ea6\u53ef\u5206\u79bb\u5377\u79ef\u4e0b\u91c7\u6837\uff088x depthwise-separable convolutional downsampling\uff09<\/strong>\uff0c\u4ee5\u63d0\u9ad8\u8ba1\u7b97\u6548\u7387\u3002<\/p>\n\n\n\n<p><strong>TDT\uff08Token-and-Duration Transducer\uff09<\/strong> \u662f\u5bf9\u4f20\u7edf Transducer \u7684\u4e00\u79cd\u6cdb\u5316\u65b9\u5f0f\uff0c\u5b83\u5c06 <strong>\u201c\u97f3\u7d20\uff08token\uff09\u201d\u4e0e\u201c\u6301\u7eed\u65f6\u95f4\uff08duration\uff09\u201d\u7684\u9884\u6d4b\u8fc7\u7a0b\u89e3\u8026<\/strong>\u3002\u4e0e\u4f20\u7edf Transducer \u5728\u63a8\u7406\u9636\u6bb5\u4ea7\u751f\u5927\u91cf\u7a7a\u767d\uff08blank\uff09\u8f93\u51fa\u4e0d\u540c\uff0cTDT \u6a21\u578b\u53ef\u4ee5\u901a\u8fc7\u6301\u7eed\u65f6\u95f4\u9884\u6d4b\u8df3\u8fc7\u5927\u591a\u6570 blank\uff08\u4f8b\u5982\u672c\u6a21\u578b parakeet-tdt-1.1b \u6700\u591a\u53ef\u8df3\u8fc7 4 \u5e27\uff09\uff0c\u4ece\u800c\u5927\u5e45\u63d0\u5347\u63a8\u7406\u901f\u5ea6\u3002\u5173\u4e8e TDT \u7684\u8be6\u7ec6\u5185\u5bb9\uff0c\u8bf7\u53c2\u89c1\u6587\u7ae0\uff1aEfficient Sequence Transduction by Jointly Predicting Tokens and Durations\u3002<\/p>\n\n\n\n<p>The training dataset consists of private subset with 40K hours of English speech plus 24K hours from the following public datasets:<\/p>\n\n\n\n<ul><li>Librispeech 960 hours of English speech<\/li><li>Fisher Corpus<\/li><li>Switchboard-1 Dataset<\/li><li>WSJ-0 and WSJ-1<\/li><li>National Speech Corpus (Part 1, Part 6)<\/li><li>VCTK<\/li><li>VoxPopuli (EN)<\/li><li>Europarl-ASR (EN)<\/li><li>Multilingual Librispeech (MLS EN) &#8211; 2,000 hour subset<\/li><li>Mozilla Common Voice (v7.0)<\/li><li>People&#8217;s Speech &#8211; 12,000 hour subset<\/li><\/ul>\n\n\n\n<p>\u81ea\u52a8\u8bed\u97f3\u8bc6\u522b\uff08ASR\uff09\u6a21\u578b\u7684\u6027\u80fd\u901a\u5e38\u901a\u8fc7\u8bcd\u9519\u8bef\u7387\uff08Word Error Rate, WER\uff09\u6765\u8861\u91cf\u3002\u7531\u4e8e\u8be5\u6570\u636e\u96c6\u5728\u591a\u4e2a\u9886\u57df\u4e0a\u8fdb\u884c\u4e86\u8bad\u7ec3\uff0c\u5e76\u4e14\u5305\u542b\u4e86\u66f4\u5927\u89c4\u6a21\u7684\u8bed\u6599\u5e93\uff0c\u56e0\u6b64\u5728\u901a\u7528\u97f3\u9891\u8f6c\u5199\u4efb\u52a1\u4e2d\u901a\u5e38\u8868\u73b0\u66f4\u597d\u3002<\/p>\n\n\n\n<p>\u4e0b\u8868\u603b\u7ed3\u4e86\u672c\u96c6\u5408\u4e2d\u5404\u53ef\u7528\u6a21\u578b\u5728\u4f7f\u7528<strong>Transducer \u89e3\u7801\u5668<\/strong>\u4e0b\u7684\u6027\u80fd\u8868\u73b0\u3002\u6240\u6709 ASR \u6a21\u578b\u7684\u6027\u80fd\u5747\u4ee5\u8d2a\u5a6a\u89e3\u7801\uff08greedy decoding\uff09\u65b9\u5f0f\u8ba1\u7b97\u7684 <strong>\u8bcd\u9519\u8bef\u7387\uff08WER%\uff09<\/strong> \u8fdb\u884c\u62a5\u544a\u3002<\/p>\n\n\n\n<figure class=\"wp-block-table\"><table><thead><tr><th>\u6a21\u578b<\/th><th><strong>Tokenizer<\/strong><\/th><th><strong>Vocabulary Size<\/strong><\/th><th><strong>AMI<\/strong><\/th><th><strong>Earnings-22<\/strong><\/th><th><strong>Giga Speech<\/strong><\/th><th><strong>LS test-clean<\/strong><\/th><th><strong>SPGI Speech<\/strong><\/th><th><strong>TEDLIUM-v3<\/strong><\/th><th><strong>Vox Populi<\/strong><\/th><th><strong>Common Voice<\/strong><\/th><\/tr><\/thead><tbody><tr><td>\u6307\u6807<\/td><td>SentencePiece Unigram<\/td><td>1024<\/td><td>15.90<\/td><td>14.65<\/td><td>9.55<\/td><td>1.39<\/td><td>2.62<\/td><td>3.42<\/td><td>3.56<\/td><td>5.48<\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<h4>\u6838\u5fc3\u4f18\u52bf<\/h4>\n\n\n\n<ul><li>\u2022&nbsp;<strong>\u6781\u81f4\u8f6c\u5f55\u6548\u7387<\/strong>\uff1a60 \u5206\u949f\u97f3\u9891\u4ec5\u9700 1 \u79d2\u5185\u5b8c\u6210\u8f6c\u5f55\uff08A100 \u63a8\u7406\uff09<\/li><li>\u2022&nbsp;<strong>OpenASR \u699c\u9996\u8868\u73b0<\/strong>\uff1a\u8d85\u8d8a Whisper\u3001Conformer\u3001Wav2Vec \u7b49\u4e3b\u6d41\u95ed\u6e90\u6a21\u578b<\/li><li>\u2022&nbsp;<strong>\u6781\u5c0f\u53c2\u6570\u91cf<\/strong>\uff1a\u4ec5 0.6B\uff08\u8f7b\u91cf\u7ea7\uff0c\u9002\u5408\u8fb9\u7f18\u8bbe\u5907\uff09<\/li><li>\u2022&nbsp;<strong>\u9ad8\u7cbe\u5ea6<\/strong>\uff1a\u5e73\u5747 WER 6.05%\uff08Hugging Face Open ASR \u699c\u5355\uff09\uff0c\u4f18\u4e8e Whisper-large-v3<\/li><li>\u2022&nbsp;<strong>\u9ad8\u9c81\u68d2\u6027<\/strong>\uff1a\u591a\u8bed\u901f\u3001\u591a\u53e3\u97f3\u3001\u591a\u5f55\u97f3\u73af\u5883\u4e0b\u8868\u73b0\u7a33\u5b9a\uff08\u82f1\u6587\uff09<\/li><\/ul>\n\n\n\n<h4>\u5e94\u7528\u573a\u666f\u63a8\u8350<\/h4>\n\n\n\n<ul><li>\u2022 \u5b9e\u65f6\u4f1a\u8bae\u8f6c\u5199<\/li><li>\u2022 \u624b\u673a\/\u8bbe\u5907\u7aef\u8bed\u97f3\u52a9\u624b<\/li><li>\u2022 \u89c6\u9891\u5b57\u5e55\u751f\u6210<\/li><li>\u2022 \u5927\u6a21\u578b\u97f3\u9891\u8f93\u5165\u9884\u5904\u7406\u5668<\/li><li>\u2022 \u6559\u80b2\/\u8bfe\u7a0b\u8f6c\u5f55\u7cfb\u7edf<\/li><\/ul>\n\n\n\n<h4>\u6280\u672f\u6784\u5efa\u8bf4\u660e<\/h4>\n\n\n\n<ul><li>\u2022 \u67b6\u6784\uff1aTDT\uff08Time-Depth Transformer\uff09\uff0c\u4e13\u6ce8\u4e8e\u65f6\u95f4\u7ef4\u5ea6\u5efa\u6a21<\/li><li>\u2022 \u6570\u636e\uff1a\u82f1\u4f1f\u8fbe\u81ea\u5efa + \u516c\u5171\u8bed\u97f3\u6570\u636e\u96c6\u5927\u89c4\u6a21\u8bad\u7ec3<\/li><li>\u2022 \u63a8\u7406\u5f15\u64ce\u4f18\u5316\uff1a\u652f\u6301 TensorRT \/ ONNX Runtime \u7b49\u9ad8\u6027\u80fd\u90e8\u7f72\u65b9\u6848<\/li><\/ul>\n","protected":false},"excerpt":{"rendered":"<p>https:\/\/huggingface.co\/nvidia\/parakeet-tdt-1.1b parakee &hellip; <a href=\"http:\/\/139.9.1.231\/index.php\/2025\/05\/11\/asr-en-parakeet-tdt-0-6b-v2\/\" class=\"more-link\">\u7ee7\u7eed\u9605\u8bfb<span class=\"screen-reader-text\">\u82f1\u6587\u8bed\u97f3\u8bc6\u522b\u6a21\u578b\uff1aParakeet TDT 0.6B V2<\/span><\/a><\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[4,38,34],"tags":[],"_links":{"self":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/26246"}],"collection":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/comments?post=26246"}],"version-history":[{"count":20,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/26246\/revisions"}],"predecessor-version":[{"id":26267,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/posts\/26246\/revisions\/26267"}],"wp:attachment":[{"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/media?parent=26246"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/categories?post=26246"},{"taxonomy":"post_tag","embeddable":true,"href":"http:\/\/139.9.1.231\/index.php\/wp-json\/wp\/v2\/tags?post=26246"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}